Improved Anisotropic Diffusion Model for Image Enhancement

based on edge preserving approach using Self Organizing and

Bayesian Inference methods in various discretization schemes

D. A. Karras1, G. Β. Mertzios2
1
Chalkis Institute of Technology, Dept. Automation and Hellenic Open University,
Rodu 2, Ano Iliupolis, Athens 16342, Greece, e-mails: dakarras@teihal.gr, dakarras@ieee.org,
dakarras@usa.net
2
Fakultat fur Mathematik und Informatik, Technische Universitat Munchen,
Boltzmannstr 3, D-85748 Garching Munchen, Germany, Mertzios@in.tum.de

Abstract – A novel approach is presented in this paper for improving anisotropic diffusion PDE models, based on
Perona-Malik equation, proposing a solution, from an engineering perspective, to adaptively estimate the
parameters of the regularizing function involved in this equation. The goal of such a new adaptive diffusion scheme
is to better preserve edges when the anisotropic diffusion PDE models are applied to image enhancement tasks. The
proposed adaptive parameter estimation in the anisotropic diffusion PDE model involves self-organizing maps and
Bayesian inference in order to define edge probabilities more accurately. The proposed modifications attempt to
capture not only simple edges but, also, the more difficult textural edges and incorporate their probability in the
anisotropic diffusion model. In the context of PDE models application to image processing such adaptive schemes
are closely related to the discrete image representation problem and the investigation of more suitable discretization
algorithms using constraints derived from image processing theory. The proposed adaptive anisotropic diffusion
model illustrates these concepts when it is numerically approximated by various discretization schemes in a database
of Magnetic Resonance Images (MRI), where it is shown to be quite efficient in image filtering and restoration
applications.

Keywords - Anisotropic Diffusion PDEs, PDE Discretization, Image Impainting, Bayesian Inference, Self-
Organizing Maps, Edge-stopping diffusion

I. INTRODUCTION

While various PDE models are in discussion since ten years and are widely applied nowadays in image
processing and computer vision tasks [1], including restoration, filtering, segmentation and object tracking the
perspective adopted in the majority of the relevant reports is the view of applied mathematician, attempting to
prove existence theorems and devise exact numerical methods for solving them. Unfortunately, such solutions
are exact for the continuous PDEs but due to the discrete approximations involved in image processing the
results yielded might be unsatisfactory. There is need to investigate, from an engineering perspective, how to
incorporate sophisticated image processing algorithms in the discretization schemes used in the numerical
analysis of continuous PDE models when applied to image processing tasks in order to achieve better results.
There are only a few reports in the literature presenting the computational perspective of the available
PDE models with respect to image processing models suitably incorporated in these PDE methods. The
majority of efforts are focused on devising new continuous PDE models and analyzing their solutions in
general mathematical spaces, having as application example image datasets, but they pay less attention to
enhance the results of existing PDE models by incorporating sophisticated image processing algorithms in
their numerical solution. In order to efficiently apply PDE models in image processing it is important to
merge the two fields, namely, numerical analysis of PDE models and image processing techniques. The
goal of this paper is to illustrate this concept in the application of anisotropic diffusion PDE models in
image enhancement tasks. It is shown that such PDE models could be improved if sophisticated edge-
stopping diffusion schemes were incorporated in the models and their numerical solutions.
In order to achieve better image enhancement results using anisotropic diffusion PDE models it is
necessary to preserve edges [1] and especially textural edges. To this end, we present a modification of the
anisotropic diffusion model of Perona-Malik [4] by adaptively estimating the parameters introduced in the
regularization function involved. Such an adaptive estimation is based on Bayesian inference and self-
organization schemes. The better results achieved, which are illustrated in the experimental section of the
paper with respect to image filtering and restoration applications, lead to the idea that many other PDE
models for image impainting tasks [2], attempting to prevent edge blurring, would be enhanced by
integrating suitable image processing techniques in their formalism and their discretization and numerical
approximation, through adaptive estimation of model parameters. It is, therefore, important to critically
overview, from such an engineering perspective, the principles of these PDE models as well as their
associated model parameters. Our overview will be naturally focused, however, on anisotropic diffusion
PDEs. The proposed adaptive anisotropic diffusion model is numerically approximated using different
discretization schemes in order to investigate their suitability. More specifically, the Finite Difference
method (FDM) and the Radial Basis Function (RBF) discretizing scheme are employed to solve the
proposed PDE models.
The organization of this paper is as follows. In section II a critical overview of the principles of PDE
models applied to image impainting tasks involving edge preserving schemes is presented with emphasis
in anisotropic diffusion models for image enhancement. In section III special attention is paid to the
discretization schemes and numerical approximations of such PDE models with emphasis in Finite
Difference and Radial Basis Functions discretization methods that are used in the experimental section of
the paper. Section IV describes the new adaptive anisotropic diffusion model and section V illustrates the
results achieved by its application to a large set of MRI images. Finally, section VI concludes the paper.

II. IMAGE IMPAINTING PDE MODELS PREVENTING EDGE BLURRING

Partial Differential Equations (PDEs) have led to an entire new field in image processing and computer
vision. They offer several advantages:
• Reinterpretation of traditional techniques under a novel unifying framework. This includes many
known techniques such as convolution, filtering and morphological operations of dilation/erosion.
• More invariances could be offered with respect to classical techniques
• Better mathematical modeling, connection with physical phenomena and better approximation to the
geometry (Euclidean or generalized) of the problem. Guaranteed mathematical results with respect to
well-posedness are available, such as proving that the numerical algorithms involved are stable.
• Shape recognition, structure preserving filtering, object segmentation could be performed within a
new more intuitive framework
Regarding image impainting [2], there is a variety of available PDE models proposed mainly to smooth
and denoise images. In the last two decades, the use of nonlinear PDEs for image smoothing and
denoising has met with tremendous success [1]. Before nonlinear PDEs were introduced, images were
denoised by linear filtering, which is equivalent to using the noisy image as an initial condition for the
heat equation. Although this method removes high frequency noise, it also badly blurs edges. To prevent
blurring, a number of authors suggested using a nonlinear diffusion equation or a variational PDE model.
Since the present paper investigates a novel way on preserving edges when the anisotropic diffusion PDE
models are applied, by integrating an adaptive strategy for model parameter estimation, it is important to
summarize both diffusion and variational approaches.
Among the equations belonging in the first category the most famous example is the Perona-Malik
equation [3,4]. On the other hand, among the most famous examples of PDEs belonging in the second
category is the variational Mumford-Shah model [1,3,5] as well as the Total Variation (TV) model [6,7].
Although effective, the methods produce piecewise constant images, often giving “blocky” results. The
Perona-Malik equation for instance behaves as a backwards heat equation and instantly creates jumps (i.e.
shocks) in unpredictable locations [8]. Such results occur either because of the PDE model which might be
not representative of the image dynamical system or because of the discretization and numerical
approximation schemes involved.
In an attempt to improve upon the piecewise constant images resulting from second order image
diffusions, fourth order diffusions have been suggested for image denoising mainly. Examples include the
‘Low Curvature Image Simplifier’ (LCIS) equation of Tumblin and Turk [8,9] as well as similar higher
order PDE models [10]. Other mainstream such attempts include the imposing of constraints in the
diffusion PDE models [11]. The herein proposed adaptive anisotropic diffusion scheme could be viewed
as belonging in such a line of research.
The main trend is, therefore, investigation of nonlinear PDE models, although it might be difficult or
even impossible to analyze. Moreover, the majority and the most useful image analysis techniques are
nonlinear, which is due to the inability of linear systems to successfully model important problems. The
most known vision problem modeled via PDEs is that of multiscale analysis, which is a useful and often
required framework for many tasks such as feature/object detection, motion detection, stereo and multi-
band frequency analysis or even image enhancement as it is the case of this study.
 Consider a multiscale operator Tt mapping an input image f to an output image Tt ( f ) , which results

from the interaction of f with some kernel function dependent on a continuous scale parameter t ≥ 0 ,

i.e. Tt ( f )( x, y ) = u ( x, y, t ) . The scale-space function u ( x, y, t ) holds all the history of transforming

f though all the scales and van be viewed as the output of Tt at any fixed scale t . The evolution of u in
scale-space as a continuous dynamical system can be modeled by evolution PDEs of the
type ut ( x, y , t ) = function (u xx , u xy , u yy , u x , u y , u , x, y , t ) and u can be viewed as the solution of the PDE

with initial condition u ( x, y, 0) = f ( x, y ) . This multiscale analysis offers the unified framework for
applying diffusion models to image processing tasks and, more specifically, to image enhancement. The
most important PDE models and their variations for solving image enhancement problems are below
summarized, taking into account their strategies in preventing edge blurring.

A. Linear heat-diffusion PDE.

The most investigated PDE method for smoothing images is to apply a linear diffusion process for
modeling Gaussian scale-space [12]. The convolution of an image with a Gaussian function of increasing
variance is equivalent from a physical point of view with linear diffusion filtering. The connection of
Gaussian convolution and linear diffusion filtering extends its limits to multiscale analysis. When there is
not clear in advance what is the right scale it is desirable to have an image representation at multiple
scales. Diffusion could be thought as a physical process that equilibrates concentration differences without
creating or destroying mass. The equilibration property is expressed by Fick’s law:

j = − D ⋅ ∇u (1)

which states that a concentration gradient ∇u causes a flux j , aiming to compensate the gradient. The

relation between ∇u and j is described by the diffusion tensor D , a positive definite symmetric matrix.
The case where ∇u and j are parallel is called isotropic. Then the diffusion tensor may be replaced by a
positive scalar-valued diffusivity g .In the general anisotropic case, ∇u and j are not parallel.
The observation that diffusion only transports mass without destroying it or creating new mass is
expressed by the continuity equation:
∂ t u = − div ( j ) (2)
where t denotes the time. By connecting Fick’s law with the continuity equation we end up with the
diffusion equation:
∂ t u = div ( D ⋅ ∇u ) (3)
 Equation (3) appears in many physical transport processes. In image processing we may associate the
concentration with the gray value at a certain location.

B. Anisotropic Diffusion PDE Models

Since linear filtering causes edge blurring and linear shifting, the development of anisotropic nonlinear
diffusion PDEs for multiscale directional image smoothing and edge detection was motivated. Perona and
Malik ([4]) proposed a nonlinear diffusion method for avoiding the blurring and localization problems of
linear diffusion filtering (hence Tt nonlinear). This Perona-Malik scheme appears to be the finite

differences discretization of a nonlinear PDE not followed by a theory of well-posedness. It was known
that, despite its success at its intended purpose the scheme is very sensitive to the presence of noise and
the choice of parameters such as the resolution of the digital image - a fact intimately connected with the
lack of a continuum PDE theory. The work of Lions et al. [13] replaced the Perona-Malik scheme with
one that has all the desirable characteristics of the original, as well as a rigorously established continuum
limit. A scale – space is an image representation at a continuum of scales, embedding the image into a
family of gradually simplified versions of it [13]. The practical implication is much more stable behavior
with respect to the presence of noise and different resolutions.
Computationally, solving the modified Perona-Malik anisotropic diffusion equation, mainly following
Rothe's approximation in time and finite element method in space, involves the PDE
∂u / ∂t − div( g (| ∇Gσ ⊗ u |)∇u ) = f (u 0 − u ),....with....u ( x, y,0) = u 0 ( x, y ) (4)

together with zero Neumann boundary conditions and initial condition representing the processed image.
Here, g(s) tends to 0 for s tending to infinity. It causes the selective smoothing of the image regions and
keeping of the edges on which the 'gaussian gradient' is large (Gσ is smoothing kernel of the convolution
denoted by the operator ⊗ in equation (4)). Such image analysis is included in the so-called nonlinear
scale space theory.

C. Variational PDE Models

The variational approach to the image denoising problem seeks to exhibit the "restored" image as the
minimizer of a functional defined over the space of all images. The first task is clearly to decide which
space of functions to take images from. For example, Sobolev spaces are ill suited for this purpose since
their elements cannot have discontinuities. Such discontinuities need to be allowed because one of the
most important features of images, namely "edges" correspond squarely to this type of behavior.
A variational approach has been proposed [5] for the solution of the image segmentation problem,
where the segmentation is obtained by finding the minimizer of an energy function, given an original
image. The correct space of functions for energy minimization turns out to be a subset of functions of
bounded variation. The Mumford-Shah model is a non-typical variational problem, whose analysis led to a
wealth of new mathematics. Numerical implementation of the Mumford-Shah model has also been a
subject of intense mathematical research. The energy is very difficult to handle since it requires
minimization over subsets. The work of L. Ambrosio and V. M. Tortorelli [1] has rigorously shown how
to approximate it in the sense of Gamma convergence by elliptic functionals. In a different vein, the work
of T. Chan and L. Vese [1] has shown how the level set method of S. Osher and J. Sethian can be
effectively utilized in the minimization of these types of energies.
Another successful example of the variational and PDE method is the Total Variation (TV)
minimization [6]. An improved version of the latter technique that is based on the Connectivity Principle
is the Curvature Driven Diffusions (CCD) impainting scheme [3,14].

III. METHODS FOR ANISOTROPIC DIFFUSION PDEs NUMERICAL SOLUTION AND

DISCRETIZATION

The numerical solution of partial differential equations (PDEs) has been dominated by finite difference
methods (FDM), finite element methods (FEM) and finite volume methods (FVM). These methods can be
derived from the assumptions of the local interpolation schemes. These methods require a mesh to support
the localized approximations; the construction of a mesh in three or more dimensions is a non-trivial
problem. Typically with these methods only the function is continuous across meshes, but not its partial
derivatives.
In practice, only low order approximations are used because of the polynomial snaking problem. While
higher order schemes are necessary for more accurate approximations of the spatial derivatives, they are
not sufficient without monotonicity constraints. Because of the low order schemes typically employed, the
spatial truncation errors can only be controlled by using progressively smaller meshes. The mesh spacing,
h, must be sufficiently fine to capture the functions of the partial derivative behavior and to avoid
unnecessarily large amounts of numerical artifacts contaminating the solution. Spectral methods while
offering very high order spatial schemes typically depend upon tensor product grids in higher dimensions
[15].
The last decade the idea of using mesh-free methods for the numerical solution of PDEs has received
much attention [16]-[17] including RBF based methods and wavelets [18]-[19]. In this paper we compare
the solutions offered by the proposed adaptive anisotropic diffusion scheme when the relevant PDEs are
numerically solved with the FDM method as well as with the RBF mesh-free technique.
A. Solving PDEs with Radial Basis Functions (RBF)

The idea of solving numerically PDEs based on Radial Basis Functions (RBFs) mostly has dealt with
elliptic problems, although some efforts have been made to solve time dependent parabolic or hyperbolic
problems. RBFs were first introduced to scattered data fitting and to numerical solution of PDEs [15].
This was done in the form of globally supported RBFs and specifically of multiquadrics (MQ)
2 2
φ (r ) = r 2 + c 2 of thin plate splines φ (r ) = r 2 log r , or Gaussians φ (r ) = e − c r , where r = x with
c ≠ 0 is a parameter.
For the solution of the scattered data fitting problem [20] an RBF based expansion of the form

∑c φ ( x − x ),
n
s ( x) = j j 2 x∈Ω ⊆ d
j =1

is involved and then, the coefficients c j are determined by satisfying the interpolation conditions

s ( xi ) = f ( xi ), i = 1, 2,..., n
Where, f is a known function that generates the data to be fitted. There exists a trade-off principle,
which says that the spectral convergence is achieved at the cost of instability.
Galperin and Zheng [21] argue that all collocation methods are intrinsically ill conditioned. Ill-posed
and badly formulated problems can possess-equivalent solutions that represent physical reality despite the
mathematical nonexistence of an exact solution. Only Galperin, Pan, and Zheng [22] have used global
optimization on a few limited problems, with extra-ordinary results.
Although it is clear that the numerical solutions of PDE, ODE, integral, and integro-differential
equations would greatly benefit from the global optimization, the major implementation impediment is the
lack of robust multi-parameter global optimization software. Unfortunately, gradient based methods are
ill-conditioned, and converge rapidly only under certain restricted conditions. In addition, gradient
methods pose the risk of being trapped in a local minimum, rather than in the global minimum. Ferrari and
Galperin [23] have published a software package of a fast one-dimensional adaptive cubic algorithm. It is
hopeful that fast multi-dimensional global optimization software packages would be developed soon.
Therefore there are serious limitations to the applicability of global methods and for large dimension
problems the solution should employ localization schemes ([24]). The localization of the basis functions
leads to locally (compactly) supported RBFs. One of the most popular compactly supported RBFs has
been proposed by Wendland ([24]) in order to use compactly supported RBFs in the context of scattered
data fitting or to solve PDEs, an hierarchical strategy has been developed leading to a multilevel
algorithm, in which residuals are fitted iteratively and are used to update the solution. Then, the
computational complexity is held proportional to the number of points used at the computational mesh,
while achieving at least linear convergence. The addition of an approximate smoothing is also needed that
ensures super-linear convergence in an entire scale of Sobolev (or more generally Besov) spaces ([24]).

IV. THE PROPOSED ADAPTIVE ANISOTROPIC DIFFUSION EDGE PRESERVING PDE

MODEL

We consider the nonlinear PDE equation of Perona and Malik [4],

∂u / ∂t = div( g (| ∇u |)∇u ),....with....u ( x, y,0) = u 0 ( x, y ) (5)

as well as the modified anisotropic diffusion model, proposed in [13] in order to remove the
inconsistencies of the Perona-Malik model,

∂u / ∂t = div( g (| ∇Gσ ⊗ u |)∇u ),....with....u ( x, y,0) = u 0 ( x, y ) (6)

The latter anisotropic diffusion model is a special case of equation (4) and it is the model herein
investigated and improved by incorporating an adaptive scheme for estimating its parameters. In the above
equations (5) and (6) g is a smooth non-increasing function with g(0)=1, g(s)≥0 and lim g(s) =0 for s->∞.
These properties of g function cause, that the diffusion is a conditional process such that when | ∇u | is

large, then, the diffusion is small and the position of edges is kept. If, on the other hand, | ∇u | is small,
then, the diffusion is large and the image in the neighborhood of the point (x,y) is smoothed. The g
functions most widely used are

g ( s ) = 1 /(1 + s 2 / λ2 ),...and ...g ( s ) = 1 / e λs (7)

The former of these functions is employed in this paper, that is, g ( s ) = 1 /(1 + s 2 / λ2 ) in the proposed
adaptive anisotropic diffusion model. The parameter of this model is, therefore, the coefficient λ, and an
adaptive scheme is herein suggested for its estimation in order to improve the corresponding anisotropic
diffusion model in terms of better preserving edges.

In the proposed adaptive scheme the coefficient λ(x,y) is adaptive and could be defined as the inverse
probability that the point (x,y) belongs to an image u edge. If such a probability p(x,y) is small and
p(x,y)->0 then, function g(s) takes on its largest value and g(s) -> 1. Therefore, the diffusion increases. On
the other hand, if p(x,y)->1 then, function g(s) takes on smaller values and the diffusion is small. That is, a
definition of λ(x,y) = 1/p(x,y) is reasonable in terms of achieving an edge preserving diffusion model.
Such an enhancement is important for the anisotropic diffusion model since | ∇u | values are only noisy
estimates of edge existence, being the only guidance for the diffusion process. Therefore, the basic idea
underlying the proposed adaptive scheme is that edge preservation could be enhanced by incorporating
more guidance in the diffusion process with regards to edge detection.

In order to estimate p(x,y) in the proposed adaptive anisotropic diffusion model we involve a two stage
methodology. Self-Organizing Feature Map (SOM) of Kohonen [25] is the first step while the Bayesian
inference procedure is the second stage performed in order to fine tune SOM probability estimates.

A) First, regarding SOM application for estimating edge probabilities p(x,y), we could consider
image u(x,y,t) for each iteration t of the anisotropic diffusion PDE solution process, and follow
the next steps.

1) The N x N image u(x,y,t) is raster scanned by M x M sliding windows having centers the
points (x,y). For each such point (x,y) its gradient | ∇u | ( x , y ) is estimated to become an input

characteristic of the SOM network. The output SOM map consists of K x K processing
elements that is, K x K codebook vectors of M x M dimensions each. The goal of SOM is to
cluster | ∇u | ( x , y ) space in edge and non edge points providing a measure of such a
probability. SOM is known as a topology preserving map [25], having the capability to cluster
input space keeping its probability distribution. It is not known, however, how to extract such
aposteriori probability distribution measures from winning codebook vectors. Aposteriori
Probability estimation is known only in Multilayer Perceptrons (MLP) neural networks. In the
followings we demonstrate how a solution could be provided for such a SOM network too.
2) After the SOM map convergence process is finished following the known Kohonen algorithm
[25], its codebook vectors encode the topological space of the | ∇u | ( x , y ) domain by

preserving input vectors probability distribution and are estimated as the associated with the
SOM map weight vectors. Let’s Cb1(I1x1,….,Ix,y, …IMxM), Cb2(I1x1,….,Ix,y, …IMxM),.
CbKxK(I1x1,….,Ix,y, …IMxM) stand for these codebook vectors and (I1x1,….,Ix,y, …IMxM) stand for
the inputs of the SOM network, belonging to the M x M sliding window on the | ∇u | ( x , y )

domain. (x,y) is the central point of this window with input value | ∇u | ( x , y ) .

3) The SOM weight Cbwin(Ix,y) corresponding to the winning element/codebook Cbwin, when an
input vector (I1x1,….,Ix,y, …IMxM) is presented to the network, is a measure of | ∇u | ( x , y ) in
 terms of topology preserving clustering, It might be considered as the best representative of
| ∇u | ( x , y ) , with its noisy characteristics removed. This follows from the encoding properties
of SOM [25] codebook vectors and it is important in order to derive a less noisy estimate of
| ∇u | ( x , y ) , since such an estimate only offers the capability to identify edge areas. Therefore,
if |Cbwin(Ix,y)| is large with respect to all other Cbr(Ix,y) codebook vectors weights, then, it is
more probable that (x,y) belongs to an edge area.
4) Therefore, if we consider

p0(x,y)= [Cbr(Ix,y) – min { Cbr(Ix,y)}]/ [max { Cbr(Ix,y)}– min { Cbr(Ix,y)}] (8)

then, p0(x,y) could be a measure of the probability that (x,y) belongs to an edge area of image
u(x,y). That is, p0(x,y) over the image space u(x,y,t) is an initial estimate of the probability
distribution p(x,y) of edge areas in the image space u(x,y). We have demonstrated, by the
above described scheme, how a SOM network could provide a measure of the probability
distribution of its input space.

B) Regarding, the second stage of the proposed adaptive anisotropic diffusion scheme, it is based on
Bayesian inference [25] The proposed algorithm amounts to minimizing the following objective
function, and it is similar to the one proposed in [26]

| P0- P |2/(2σ2) + (3/2) ∑x,y log { α2 + (x Δxy)2 + (yΔxy)2 } (9)

with regards to P , which is the unknown probability distribution to be reconstructed from the
initial estimates p0(x,y) given in P0. The first term comes from the likelihood term and the second
one from the prior knowledge term of the well known Bayesian formulation. The second term
symbols arise from the imposed 2D Lorentzian prior knowledge. xΔxy and yΔxy are the probability
differences in the x- and y- directions respectively and α is a Lorentz distribution-width
parameter. If we assume that PP(p) is the prior, which imposes prior knowledge conditions about
the edge probability distribution on the reconstruction algorithm, then, the second term of (9)
comes as follows.
The starting point is that PP(p) could be obviously expanded into PP(p)=PP(p0,0) PP(p1,0| p0,0)
PP(p2,0| p0,0, p1,0 )… If, now, it is assumed that the probability px,y depends only on its left neighbor

( px-1,y ), then the previous PP(p) expansion takes on the form PP(p) = ∏(x,y) PP(px,y| px-1,y),

provided that the boundaries are ignored. Next, we assume that PP(px,y| px-1,y) is a function only of
 the difference between the corresponding probabilities. This difference is written down as xΔxy =
x
px,y - px-1,y. It has been shown that the probability density function of Δxy is Lorentzian shaped
(see [26]). These assumptions and calculations lead directly to compute the prior knowledge in the
Bayesian reconstruction approach as in (9). It should be noted that the herein proposed objective
function of (9) is minimized using the conjugate gradients optimization methodology

Therefore, during the first stage of the proposed scheme we obtain the initial estimates p0(x,y) of the edge
areas probability distribution and during the second stage we reconstruct the edge probability distribution
space p(x,y). In the sequel, λ(x,y) could be defined as λ(x,y)= 1/p(x,y), as already mentioned. This final
step concludes the proposed adaptive parameter estimation scheme as a novel modification of the
anisotropic diffusion model under investigation.

V. EXPERIMENTAL EVALUATION OF THE PROPOSED ADAPTIVE DIFFUSION MODEL

An experimental study has been conducted in order to evaluate the proposed adaptive anisotropic
diffusion PDE model described in section IV as compared to the standard modified Perona-Malik
anisotropic diffusion PDE model [13] of equation (6), and numerically solved using the discretization
schemes outlined in section III above, namely, the Finite Difference Method (involving the Crank-
Nicholson scheme) as well as the RBF meshfree technique (involving the scheme proposed by Wendland
([24])). All simulations have been performed in MATLAB version 6.5 system.
The methods involved have been applied to an MRI image database which has been downloaded from
the Internet, namely, the Whole Brain Atlas http://www.med.harvard.edu/AANLIB/ home.html (copyright
© 1995-1999 Keith A. Johnson and J. Alex Becker). These images have 256 by 256 dimensions and
present difficulties for image analysis due to their fine edges. The sliding window used by the SOM
network in the proposed adaptive scheme is 5 x 5 dimensions, while the neighborhood used in all
numerical approximations is the usual cross-neighborhood of 4 points as shown next.

Concerning the measures involved to quantitatively compare the performance of the various models in
terms of image reconstruction/ filtering, we have employed the usually used Sum of Squared Errors (SSE)
between the original MRI image pixel intensities and the corresponding pixel intensities of the
reconstructed/ filtered image as well as the RMS error in dB [26].
The quantitative results obtained by the different PDE models involved are outlined in table 1.These
results show a superiority of the proposed adaptive parameter estimation anisotropic diffusion PDE model
in terms of image reconstruction performance. The results illustrated in table 1 have been derived using 45
images randomly selected from the above mentioned database. The best mean values of reconstruction
errors obtained from this sample of images, after a certain number of iterations in the numerical solution,
different for each PDE model (so as to obtain best performance for each model) are herein reported, under
different noise levels.

Discretization scheme/ numerical Proposed adaptive Standard Modified Perona

approximation method parameter estimation Malik anisotropic diffusion
anisotropic diffusion PDE model of equation (6)
PDE model

Gaussian Noise level Mean SSE Mean dB Mean SSE Mean dB

(SNR = 50 dB)

FDM (Crank-Nicholson) 2.45 E3 19.60 3.62E3 16.26

RBF Mesh-free (Wendland ([24])) 2.31 E3 20.56 3.41E3 16.97

Gaussian Noise level Mean SSE Mean dB Mean SSE Mean dB

(SNR = 20 dB)

FDM (Crank-Nicholson) 4.08 E3 14.30 4.82E3 12.22

RBF Mesh-free (Wendland ([24])) 3.52 E3 15.24 4.11E3 13.1

Gaussian Noise level Mean SSE Mean dB Mean SSE Mean dB

(SNR = 10 dB)

FDM (Crank-Nicholson) 4.96 E3 13.60 5.72E3 11.73

RBF Mesh-free (Wendland ([24])) 4.67 E3 14.1 5.53E3 12.68

Table 1. The quantitative results with regards to reconstruction/filtering performance of the various
methodologies involved, under different gaussian noise levels imposed on the original images.

The above results show that the proposed anisotropic diffusion scheme presents better performance
compared to the standard anisotropic diffusion model (modified Perona-Malik of equation (6)), in terms of
image reconstruction under different noise levels and under different discretization schemes. Moreover,
the qualitative results obtained involving the methods outlined in table 1, regarding edge preservation, are
next outlined in figure 2
Figure 1 presents an MRI image and its edges detected involving the standard Sobel algorithm, under no
noisy conditions.

Figure 1. A gray- scale MRI image and its edges detected by applying a Sobel filter to the original image.
There is no Gaussian noise applied

Figure 2 clearly illustrates the edge preservation properties of the proposed PDE model when its results
are compared with the ones obtained by applying the standard anisotropic PDE model of equation (6).

(a) (b)
 (c) (d)
Figure 2. The edges of the MRI image of figure 1, reconstructed by applying
a) The proposed adaptive anisotropic diffusion model in the original image under noise level with
SNR= 50 Db.
b) The standard modified Perona-Malik anisotropic diffusion model of equation (6) in the original
image under noise level with SNR= 50 Db.
c) The proposed adaptive anisotropic diffusion model in the original image under noise level with
SNR= 10 Db.
d) The standard modified Perona-Malik anisotropic diffusion model of equation (6) in the original
image under noise level with SNR= 10 Db.
In all above cases the FDM method of Crank-Nicholson scheme is used to numerically approximate
the PDE models, and the edges are detected using a standard Sobel filter on the best outcome of each
corresponding PDE model.

VI. CONCLUSIONS AND FUTURE TRENDS

A novel methodology is presented in this paper for improving anisotropic diffusion PDE models, based
on Perona-Malik equation, suggesting an image analysis derived scheme to adaptively estimate the
parameters of the regularizing function involved in this equation. The goal of such a new adaptive
diffusion scheme is to better preserve edges when the anisotropic diffusion PDE models are applied to
image enhancement tasks. The proposed adaptive parameter estimation in the anisotropic diffusion PDE
model involves self-organizing maps and Bayesian inference in order to define edge probabilities more
accurately. As it is illustrated in the extensive experimental study conducted, the proposed modifications
achieve capturing not only of simple edges but, also, of the more difficult textural edges. In the context of
PDE models application to image processing such adaptive schemes are closely related to the discrete
image representation problem and the investigation of more suitable discretization algorithms using
constraints derived from image processing theory. There is need to investigate, from an engineering
perspective, how to incorporate sophisticated image processing algorithms in the discretization schemes
used in the numerical analysis of continuous PDE models when applied to image processing tasks in order
to achieve better results. The proposed adaptive anisotropic diffusion model illustrates these concepts in a
set of MRI images. This is, however, a first step in the improvement of anisotropic diffusion PDE models
of Perona-Malik type. Another important step might be the application of more robust discretization
schemes, like the MD-WDF approach in numerically approximating the involved PDEs. Finally, since the
proposed scheme increases computational needs, it is important to investigate faster algorithms in
achieving similar edge preserving results.

REFERENCES
[1] T. Chan, J. Shen and L. Vese, “Variational PDE Models in Image Processing”, Notices of the AMS Journal,
Vol. 50 , n. 1, pp. 14-26, 2003.
[2] M. Bertalmio, G. Sapiro, V. Caselles, and C. Ballester. “Image impainting”. In SIG-GRAPH, pages 417 – 424,
2000.
[3] T. F. Chan and J. Shen. “Mathematical models for local nontexture impaintings”, SIAM J. Appl. Math.,
62(3):1019–1043, 2002.
[4] P. Perona and J. Malik, “Scale-space and edge detection using anisotropic diffusion”. IEEE Trans. Pattern Anal.
Machine Intell., 12:629–639, 1990.
[5] D. Mumford and J. Shah, “Optimal approximations by piecewise smooth functions and associated variational
problems”, Comm. Pure and Applied Mathematics, 42, pp. 577-684, 1989.
[6] L. Rudin, S. Osher, and E. Fatemi, “Nonlinear Total Variation (TV) based noise removal algorithms”, Physica
D, 60:259–268, 1992.
[7] S. Osher, A. Sole, and L. Vese, “Image decomposition and restoration using Total Variation (TV) minimization
and the H−1 norm”, 2002, UCLA CAM preprint 02-57, October 2002.
[8] A L. Bertozzi and J. B. Greer. “Low curvature image simplifiers: global regularity of smooth solutions and
Laplacian limiting schemes”, Comm. Pure App. Math., October 2003.
[9] J. Tumblin and G. Turk., “LCIS: A boundary hierarchy for detail-preserving contrast reduction”. In Proceedings
of the SIGGRAPH 1999, August 8-13, 1999, Los Angeles, CA USA, pages 83–90, 1999.
[10] Yu-Li You and M. Kaveh,. “Fourth-order partial differential equations for noise removal”, IEEE Trans. Im.
Process., 9(10):1723–1730, 2000.
[11] D. Tschumperle and R. Deriche, “Constrained and Unconstrained PDEs for vector image restoration”, In Ivar
Austvoll, ed. Proc. Of the 10th Scandinavian Conf. Image Analysis, pp. 153-160, Norway, 2001.
[12] J.J. Koenderink, “The structure of images”, Biol. Cybern., vol. 50, pp. 363-370, 1984.
[13] L. Alvarez, F. Guichard, P-L. Lions, J-M. Morel, “Axioms and fundamental equations in image processing”,
Arch. Rational Mech. Anal., vol.123, pp. 199-257, 1993.
[14] T. Chan and J. Shen, “Non texture impainting by Curvature Driven Diffusions (CDD)”, J. of Visual Comm. and
Image Processing, 2002.
[15] E.J. Kansa and Y.C. Hon, “A scattered data approximation scheme with applications to computational fluid
dynamics-II Solutions to parabolic, hyperbolic and elliptic partial differential equations”, Comput. Math.
Applic., vol. 19, pp. 147-161, 1990.
[16] I. Babuska and M. Melenk, “The partition of unity method”, Int. J. Numer. Methods, vol. 40, pp. 727-758, 1997.
[17] T. Belytschko, Y. Krongauz, D. Organ, M. Fleming and P. Krysl, “Meshless methods: An overview and recent
developments”, Comp. Meth. Appl. Mech. Eng., vol. 139, pp. 3-47, 1996.
[18] O.V. Vasilyev and S. Paolucci, “A fast adaptive wavelet collocation algorithm for multidimensional PDEs”, J.
Comput. Physics, vol. 138, pp. 16-56, 1997.
[19] O.V. Vasilyev, D.A. Yuen and S. Paolucci, “Solving PDEs using wavelets”, Comput. in Physics, vol. 11, pp.
429-435, 1997.
[20] R.L. Hardy, “Multiquadratic equations of topography and other irregular surfaces”, J. Geophys. Res., vol. 76, pp.
1905-1915, 1971.
[21] E.A. Galperin, and Q. Zheng, ''Solution and control of PDE via global optimization methods'', Comput. Math.
Applic., vol. 25, No. 10/11, pp.103-118, 1993.
[22] E.A. Galperin, Z. Pan and Q. Zheng, ''Application of global optimization to implicit solution of partial
differential equations'', Comput. Math. Applic., vol. 25, No. 10/11, pp. 119-124, 1993.
[23] A. Ferrari and E.A. Galperin,''Numerical experiments with one-dimensional adaptive cubic algorithm'', Comput.
Math. Applic., vol. 25, No. 10/11, pp. 47-56, 1993.
[24] G.E. Fasshauer, “Solving differential equations with radial basis functions: multilevel methods and smoothing”,
2000
[25] S. Haykin, “Neural Networks, A comprehensive foundation”, Prentice Hall, 1999
[26] M. Reczko, D.A. Karras, B.G. Mertzios, D. Graveron-Demilly, D. van Ormondt, “Improved MR image
reconstruction from sparsely sampled scans based on neural networks”, journal pattern recognition letters,
Vol.22, Number 1, pp. 35-46, January 2001